docs(research-patch): orchestration-contour prior-art verdicts — did we reinvent the wheel?

[artyhoo]

What

Answers the operator's question — «кажется мы опять переизобрели велосипед» — for the orchestration contour (/arch, /pipeline, /dispatcher, night-mode, aif-doctor, packages/runtime-bridge), compared against mattpocock/skills, obra/superpowers, and the operator-control-plane field.

Ships one research patch (195 lines) + three SSOT entries (#230/#231/#232). No code, no skill edits, no rule edits.

Closes the gap PR #1181's GLM-authored raw material explicitly left open — its own §8 states «I did not read mattpocock's actual SKILL.md file contents (only DeepWiki summaries + README)» and it issues no verdicts by design. This PR supplies the verdicts; it does not depend on #1181 merging and supersedes none of its observations.

The answer

No for the substrate — and the evidence is one-sided:

• The runtime is someone else's: packages/runtime-bridge/src/AifHandoffBackend.ts:1-2 — «adapter for the lee-to/aif-handoff runtime». Of 2 175 TypeScript lines in the bridge, zero implement an agent runtime.
• The executor loop is upstream's (SSOT chore(skills+agents): cleanup batch from 2026-05-16 skills+agents audit #64 ADOPT); .claude/skills/night-mode/SKILL.md:11 refuses to restate it.
• The ideation loop is upstream's; .claude/skills/arch/SKILL.md:36-38 invokes superpowers:brainstorming «AS IS».
• disable-model-invocation: true — mattpocock's primary progressive-disclosure mechanism — is already ours (.claude/skills/arch/SKILL.md:6), reached independently.

Yes, narrowly, for two claims in our own bodies (recorded, not fixed — see Scope):

1. .claude/skills/arch/SKILL.md:79 claims «no design-review skill exists» upstream. But skills/brainstorming/spec-document-reviewer-prompt.md ships in both 6.1.1 and 6.2.0 (orphaned — grep -rn finds no skill body referencing it), and its loop was retired at v5.0.6 after an A/B: «doubled execution time (~25 min overhead) without measurably improving plan quality… identical quality scores» (RELEASE-NOTES.md:283-290). Our §2 may still be right — its bottom-up seat fact-checks against the codebase, which upstream's document-hygiene reviewer never did — but that argument is nowhere written, and a negative-existence claim is.
2. .claude/skills/night-mode/SKILL.md:15 names SDD's 5.1.0 two-reviewer roster, retired into one task-reviewer-prompt.md (RELEASE-NOTES.md:61); :24 frames its whole-work pass as «beyond SDD's per-increment review» while SDD :391-414 has a whole-branch Final Review; :24 caps rework at ~4 rounds against SDD :320's five.

Neither is #parallel-evolution-creep yet — no duplicate mechanism was built. Both are its precondition: the wrapper no longer knows what the wrapped thing does.

Verdicts

10 rows, per capability, on both BFR §1.1 axes, each with a T16 problem-class line and a falsifier. Three retain BUILD (/arch §3 routing, night-mode overnight delta, rule-as-test layer); ADOPT/convergent for the three wrapped loops; KEEP NARROW for /arch §2 and /pipeline; one flips to WATCHLIST — SSOT #111's 2026-06-03 search missed builderz-labs/mission-control (5 884★, created 2026-02-13, i.e. before that evaluation). Its capability depth is deliberately not asserted from a description string; §5 F1 scopes the probe.

Method

SSOT consult first (18 rows cited by ID, none re-adjudicated), DeepWiki ×4 across 2 repos, WebSearch ×3 phrasings, gh api primary source, direct reads of 3 cached superpowers versions. context7 excluded per build-first-reuse-default.md §3 tooling caveat.

Scope

Proposals only in §5 (F1-F5). No edits to README.md, CLAUDE.md, .claude/rules/**, .claude/skills/**, ~/.claude/skills/** per the Artifact Ownership Contract. The two drift findings are handed to the operator as F2/F3, not fixed here.

Verification

• npx vitest run packages/core/principles/ → 35 files / 341 tests passed
• markdownlint 0 errors; link check 167 OK / 0 errors; composition 46 passed (pre-push)
• wc -l → patch 195, prior-art-evaluations.md 331 (both under the 600-line gate)

Fidelity verdict

FIDELITY: skipped — no kickoff or spec governs this PR, so there is no declared WHAT-contour for a cold fidelity auditor to check the diff against; this is an operator-asked research question answered in one patch, with no substrate declared under Provenance and no stage to gate.

§1.7 Forward-check applied

Complies with build-first-reuse-default.md §3 — all six layers run, and §1.1's two axes are applied per row rather than as one blanket verdict. Complies with ai-laziness-traps.md: T10 — §1 enumerates the whole 9-artifact population with file:line before any finding, no sampling; T3 — one inference I drew from a grep -rl (that 6.2.0 had deleted spec-document-reviewer-prompt.md) was wrong, caught by a follow-up ls, and the finding is corrected in the patch to «orphaned, not deleted» rather than shipped as-first-believed; T16 — every verdict row carries an explicit «Upstream problem class: X. Our problem class: Y. Match? evidence» line; T20 / asymmetric skepticism — the operator's prior is refused in §4.1 and upheld only narrowly in §4.2, each on file:line evidence, because unbacked agreement with the person asking is the same failure as unbacked refusal. Complies with recommendation-laziness-discipline.md §3 — a falsifier per row (docs/meta-factory/research-patches/2026-07-31-orchestration-contour-prior-art-comparison.md:118). Complies with no-paid-llm-in-ci.md (markdown only, no gate, no CI LLM call), doc-authority-hierarchy.md §2-§3 (header present at 2026-07-31-orchestration-contour-prior-art-comparison.md:3), and attention-is-not-a-mechanism.md — F4 is labelled a proposal needing its own BFR pass, and the patch does not pretend that recording drift prevents it.

§1.7 Backward-check applied

Class of this change = prior-art evaluation of an own capability. Sibling surfaces enumerated (ls docs/meta-factory/research-patches/ | grep -icE "prior-art|comparison" → 10, this file included; plus a full SSOT row scan and a rules sweep), verdicted individually — none of these is a file this diff touches:

• docs/meta-factory/research-patches/2026-05-23-meta-orchestrator-prior-art.md → SWEPT-CLEAN, not superseded: it produced /pipeline's BUILD and SSOT docs: post-1A follow-up (Class C reviewer-discipline, Q3/Q5, README absolutism) #67's REJECT; wayfinder is new evidence that reaches the same place (KEEP NARROW), so its verdict stands.
• SSOT feat(principles): principle 15 — skill paired-negative test (N2 #5, Commit B) #111 (docs/meta-factory/prior-art-evaluations.md:183) → GAP FOUND: its search missed a candidate that predates it by ~4 months. Row not rewritten (append-only); new row research(defer-reflex-detection): Stage 2 benchmark — REJECT UMBRELLA verdict #232 records the miss and F1 scopes the probe.
• SSOT chore(skills+agents): cleanup batch from 2026-05-16 skills+agents audit #64 (prior-art-evaluations.md:136) → GAP FOUND downstream, not in the row: the row's roster description is a correct historical record; the live restatement at .claude/skills/night-mode/SKILL.md:15 is stale.
• SSOT feat(skill-drift): multi-channel drift detection from D-AuditC-5 #65 → SWEPT-CLEAN: the cited Step-0 primitive re-verified live at 6.2.0/skills/using-git-worktrees/SKILL.md:21,26,33; byte-identity deliberately not claimed (diff -q shows the file changed since 6.1.1).
• SSOT docs(wave-plan): schedule channel-earliness audit (Track 2.3, parallel-safe) #179, fix(meta-orchestrator): audit round 2 — Sub-waves A/B/C/E + Sub-wave D deferred #201, chore(wave-plan): reconcile §0 to merged-PR reality (Stage 3 UNTRACKED detection) #221, docs(meta-orchestrator): output-format §4.1 — mandatory description block above each 1-liner #228 → SWEPT-CLEAN, each correct at its stated scope, none contradicted.
• .claude/rules/source-before-shape.md:20 → GAP-ADJACENT: its Origin is literally «PR feat(skill): night-mode — reusable autonomous-overnight orchestration protocol #858 shipped night-mode re-describing the loop already owned by Superpowers». Its failure mode was re-tested against both wrappers; night-mode passes the rule as written but fails on the currency of its base-description — a new variant («restates a base that has since moved») its :64 promotion criterion does not cover. Recorded, not codified: that file is maintainer-owned.
• Rules sweep grep -rlE "reinvent|parallel-evolution" .claude/rules/ → three files, not one: build-first-reuse-default.md:26,40,53,70 owns the criteria (applied, not superseded), source-before-shape.md (above), autonomous-loop-continuity.md:124 uses the tag once in passing on an unrelated surface (SWEPT-CLEAN).
• docs/meta-factory/research-patches/2026-07-26-per-role-context-shaping-raw-research.md + its 2026-07-27 addendum → ADJACENT, untouched: different question, and a parallel session owns that track this week.
• PR docs(research-patch): mattpocock/skills comparison — progressive disclosure + injection + orchestration raw material #1181 → COMPLEMENTED, not superseded: its §6 observation table survives; its §8 guess that mattpocock's repo is «personal, not a framework» is corrected by a 196 169-star measurement.

Nothing is superseded; no rule previously claimed authority over this adjudication.